NEAR’s New Token Utility and AI Economy | Illia Polosukhin

Illia Polosukhin describes NEAR AI cloud as a vertically integrated stack combining the NEAR blockchain with trusted execution environments on GPUs to deliver verifiable and private inference. Each inference produces a signed attestation containing the model hash, the prompt, the output, and the specific hardware used, making the full supply chain auditable. Polosukhin acknowledges a pragmatic trust assumption around hardware manufacturers as a current limitation of the TEE approach.

NEAR AI cloud serves open-weight models including DeepSeek, GLM, and Gemma without the system prompt filtering that providers like OpenAI and Anthropic apply. Polosukhin argues that a single OpenAI employee could theoretically modify a system prompt to manipulate the beliefs of billions of users without detection because the system is closed source, and that third-party AI routers have been shown in research to steal files, rewrite prompts, and inject malicious tool outputs in agentic contexts. The model's own training-based refusals remain independent of anything NEAR controls.

NEAR token holders can stake NEAR to receive free AI inference for themselves or their agents on the NEAR AI cloud, trading staking yield for inference capacity rather than receiving that yield as income, with inference access proportional to the amount staked. Polosukhin calls this concept universal basic AI and notes that value is generated in yield form rather than direct fee form, making it harder to equate with traditional protocol revenue. NEAR intents turned on fees in February 2026 after achieving product-market fit through 2025, with protocol fee capture fluctuating between 20 and 50 percent of NEAR emissions being captured and burned, tracked at revenue.near.org. Polosukhin expects NEAR AI to follow a similar trajectory, with staking as the first distinct primitive before fee activation. The NEAR token is described as serving three roles: a store of value granting AI capability access, block space and programmability via the blockchain, and transaction volume capture via intents.

NEAR smart contracts can call AI inference mid-transaction, pause execution, wait for the response, and then continue, with verifiable inference required when money is at stake because the entire supply chain must be verified end to end. Polosukhin uses the term autonomous businesses rather than agents to describe entities that hold NEAR on their balance sheet for always-available inference, operate continuously, access finances and execute actions in digital and real-world environments, and allow token holders to vote on mission updates while otherwise running independently. NEAR has an agent marketplace where agents can hire each other using the same intents infrastructure that insures asset swaps, providing trust, settlement, and insurance for work performed between agents.

Real deployments include Venice and Brave, which has over 100 million users, both using NEAR AI cloud for private inference, with Brave also using NEAR for end-to-end encrypted AI inference payments. The Government of Bermuda uses NEAR AI cloud for AI assistance on sensitive financial information including pensions. Abound uses it for a remittance concierge experience for Indians in the US sending money to India while also using NEAR payments infrastructure for stablecoin transactions. NEAR is described as the first L1 to implement post-quantum cryptography.

Polosukhin argues the current GPU market is opaque and dominated by multi-year contracts between cloud providers with no transparency around actual supply or demand. NEAR is implementing a compute intents market that acts as an abstraction layer allowing solvers to find compute resources globally and negotiate prices on behalf of buyers who submit underspecified requests such as GPU count and interconnect requirements, something a traditional order book cannot handle. A compute intents MVP was showcased at Nearcon, with transaction fees from compute intents flowing to the protocol. Third-party GPU providers are incentivized through NEAR emissions proportional to inference provided and cannot see user data due to the confidential architecture, though open questions around SLA and quality assurance remain unresolved. Polosukhin draws an analogy between compute and electricity, noting that unlike oil, compute cannot be stored and must be consumed immediately, which drives the dominance of long-term contracts and makes market settlement difficult.

The host notes that Chinese open-weight models like Kimi K3 are approximately 95 percent as capable as US frontier models at roughly 95 percent lower cost, and argues this benefits companies like NEAR AI that do not train models and can adopt whichever is most competitive. Polosukhin agrees and frames the broader opportunity as all economies converging into AI and blockchain. He argues the full computing stack is collapsing into an AI operating system, making traditional SaaS progressively less necessary, using Salesforce as a concrete example where AI could generate custom CRM software per organization rather than a standardized product. The host pushes back, arguing SaaS companies will not be easily commoditized because engineers working alongside AI will consistently outperform local agents lacking deep product domain expertise. Polosukhin responds that AI models are actively learning from engineers doing exactly that work and will eventually replicate it, though he acknowledges his adoption timelines are consistently too optimistic and transitions move slower than he expects.